Publicación: Segmentación de imagen ecocardiográfica mediante técnicas de aprendizaje profundo para la detección del ciclo cardíaco
Cargando...
Fecha
2025
Autores
Editor/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Educación a Distancia (UNED). E.T.S. de Ingeniería Informática
Resumen
Este Trabajo de Fin de Máster (TFM) forma parte del proyecto CAREUM y, más recientemente, del proyecto "Cribado coste-efectivo de cáncer de mama y enfermedad cardíaca reumática mediante inteligencia artificial" (PID2023-150515OB-I00). El objetivo principal de estos proyectos es desarrollar un sistema inteligente capaz de procesar imágenes ecocardiográficas, con el fin de asistir en el diagnóstico de la enfermedad cardíaca reumática.
En este contexto, el principal objetivo de mi trabajo ha sido desarrollar un método de segmentación automática de imágenes ecocardiográficas y su posterior clasificación para la detección de las fases del ciclo cardíaco.
Para comenzar, he realizado una introducción a las redes neuronales y a las redes neuronales convolucionales (CNN, por sus siglas en inglés), así como una revisión del estado del arte en el análisis de imágenes ecocardiográficas, abarcando tanto los métodos tradicionales como los basados en aprendizaje profundo (DL).
He comparado tres variantes de la arquitectura de segmentación U-Net, ajustando su complejidad y capacidad de aprendizaje para evaluar su rendimiento en la segmentación de imágenes ecocardiográficas, considerando la precisión y eficiencia computacional.
Además, se probaron tres enfoques de entrenamiento: (1) utilizando los datos originales con validación estática, (2) aplicando aumento de datos, también con validación estática, para incrementar la variabilidad, y (3) empleando los datos originales con validación cruzada para evaluar la capacidad de generalización. Estos enfoques permitieron analizar su impacto en el rendimiento de la segmentación, proporcionando información sobre mejoras en la precisión, reducción del sobreajuste y estabilidad de los modelos
Para la tarea de clasificación, se ha entrenado una red neuronal convolucional 3D (3D CNN), diseñada específicamente para mantener y procesar tanto la información espacial como la temporal de las secuencias ecocardiográficas. También se probaron otros enfoques que pueden ayudar a la detección del ciclo cardiaco, como el cálculo de la Curva de Variación de Volumen Ventricular (CVVV).
Para el modelo de segmentación, hemos obtenido en el conjunto de pruebas, una accuracy del 96.40%, un coeficiente de DICE de 96.36%, y un IoU de 93.00%. Para el de clasificación, hemos obtenido una precisión total del 95.68%, un recall del 95.37% y una medida F1 del 95.51%.
This Master's Thesis is part of the CAREUM project and, more recently, of the "Cost-effective screening for breast cancer and rheumatic heart disease using artificial intelligence" project (PID2023-150515OB-I00). The main objective of these projects is to develop an intelligent system capable of processing echocardiographic images, in order to assist in the diagnosis of rheumatic heart disease. In this context, the main objective of my work has been to develop a method for automatic segmentation of echocardiographic images and their subsequent classification for the detection of the phases of the cardiac cycle. To begin with, I have made an introduction to neural networks and convolutional neural networks (CNN), as well as a review of the state of the art in the analysis of echocardiographic images, covering both traditional methods and those based on deep learning (DL). I have compared three variants of the U-Net segmentation architecture, adjusting their complexity and learning capacity to evaluate their performance in echocardiographic image segmentation, considering accuracy and computational efficiency. In addition, three training approaches were tested: (1) using the original data with static validation, (2) applying data augmentation, also with static validation, to increase variability, and (3) using the original data with cross-validation to assess generalization ability. These approaches allowed us to analyze their impact on segmentation performance, providing insights into improvements in accuracy, overfitting reduction, and model stability. For the classification task, a 3D convolutional neural network (3D CNN) has been trained, specifically designed to maintain and process both spatial and temporal information of echocardiographic sequences. Other approaches that can help cardiac cycle detection were also tested, such as the calculation of the Ventricular Volume Variation Curve. For the segmentation model, we have obtained in the test set, an accuracy of 96.40%, a DICE coefficient of 96.36%, and an IoU of 93.00%. For the classification model, we have obtained an overall precision of 95.68%, a recall of 95.37%, and an F1 measure of 95.51%.
This Master's Thesis is part of the CAREUM project and, more recently, of the "Cost-effective screening for breast cancer and rheumatic heart disease using artificial intelligence" project (PID2023-150515OB-I00). The main objective of these projects is to develop an intelligent system capable of processing echocardiographic images, in order to assist in the diagnosis of rheumatic heart disease. In this context, the main objective of my work has been to develop a method for automatic segmentation of echocardiographic images and their subsequent classification for the detection of the phases of the cardiac cycle. To begin with, I have made an introduction to neural networks and convolutional neural networks (CNN), as well as a review of the state of the art in the analysis of echocardiographic images, covering both traditional methods and those based on deep learning (DL). I have compared three variants of the U-Net segmentation architecture, adjusting their complexity and learning capacity to evaluate their performance in echocardiographic image segmentation, considering accuracy and computational efficiency. In addition, three training approaches were tested: (1) using the original data with static validation, (2) applying data augmentation, also with static validation, to increase variability, and (3) using the original data with cross-validation to assess generalization ability. These approaches allowed us to analyze their impact on segmentation performance, providing insights into improvements in accuracy, overfitting reduction, and model stability. For the classification task, a 3D convolutional neural network (3D CNN) has been trained, specifically designed to maintain and process both spatial and temporal information of echocardiographic sequences. Other approaches that can help cardiac cycle detection were also tested, such as the calculation of the Ventricular Volume Variation Curve. For the segmentation model, we have obtained in the test set, an accuracy of 96.40%, a DICE coefficient of 96.36%, and an IoU of 93.00%. For the classification model, we have obtained an overall precision of 95.68%, a recall of 95.37%, and an F1 measure of 95.51%.
Descripción
Categorías UNESCO
Palabras clave
Citación
Mascaraque León, Aroa. Trabajo Fin de Máster: Segmentación de imagen ecocardiográfica mediante técnicas de aprendizaje profundo para la detección del ciclo cardíaco. Universidad Nacional de Educación a Distancia (UNED) 2025
Centro
E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial